Line data Source code
1 : /**
2 : * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3 : * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 : * CANN Open Software License Agreement Version 2.0 (the "License").
5 : * Please refer to the License for details. You may not use this file except in compliance with the License.
6 : * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7 : * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8 : * See LICENSE in the root of the software repository for the full text of the License.
9 : */
10 :
11 : #include "transport_remote_access.h"
12 : #include "externalinput_pub.h"
13 :
14 : namespace hccl {
15 : using namespace std;
16 : std::array<DeviceMem, MAX_MODULE_DEVICE_NUM> TransportRemoteAccess::notifyValueMem_;
17 : std::array<std::mutex, MAX_MODULE_DEVICE_NUM> TransportRemoteAccess::notifyValueMutex_;
18 : std::array<Referenced, MAX_MODULE_DEVICE_NUM> TransportRemoteAccess::instanceRef_; // 实例计数，用于释放静态资源
19 0 : TransportRemoteAccess::TransportRemoteAccess(const std::string tag, const HcclDispatcher dispatcher,
20 : const std::unique_ptr<NotifyPool> ¬ifyPool,
21 0 : const RemoteAccessPara &remoteAccessPara, const std::vector<MemRegisterAddr> &memRegistInfos, s32 deviceLogicId)
22 0 : : dispatcher_(dispatcher), notifyPool_(notifyPool), MemRegistInfos_(memRegistInfos),
23 0 : RemoteAccessPara_(remoteAccessPara), handle_(nullptr), ackNotify_(nullptr),
24 0 : access_(RA_ACCESS_LOCAL_WRITE | RA_ACCESS_REMOTE_WRITE | RA_ACCESS_REMOTE_READ),
25 0 : notifySize_(NOTIFY_BUFFER_SIZE), tag_(tag), timeout_(HCCL_LINK_TIME_OUT_S), deviceLogicId_(deviceLogicId)
26 : {
27 0 : instanceRef_[deviceLogicId_].Ref();
28 0 : }
29 :
30 0 : TransportRemoteAccess::~TransportRemoteAccess()
31 : {
32 0 : HCCL_DEBUG("~TransportRemoteAccess Enter!");
33 : HcclResult ret;
34 0 : struct MrInfoT mrInfo = {nullptr};
35 : /* 销毁本端mr */
36 0 : for (u32 idx = 0; idx < localRegMem_.size(); idx++) {
37 0 : mrInfo.addr = localRegMem_[idx];
38 0 : ret = HrtRaMrDereg(handle_, &mrInfo);
39 0 : if (ret != HCCL_SUCCESS) {
40 0 : HCCL_WARNING("errNo[0x%016llx] in TransportRemoteAccess deconstruct, mr dereg failed. ",
41 : HCCL_ERROR_CODE(ret));
42 : }
43 : }
44 :
45 0 : ackNotify_ = nullptr;
46 0 : if (handle_ != nullptr) {
47 0 : ret = HrtRaQpDestroy(handle_);
48 0 : if (ret != HCCL_SUCCESS) {
49 0 : HCCL_WARNING("errNo[0x%016llx] in TransportRemoteAccess deconstruct, qp destroy failed. ",
50 : HCCL_ERROR_CODE(ret));
51 : }
52 : }
53 0 : if (instanceRef_[deviceLogicId_].Unref() == 0) {
54 0 : std::unique_lock<std::mutex> lock(notifyValueMutex_[deviceLogicId_]);
55 0 : notifyValueMem_[deviceLogicId_].free();
56 0 : }
57 0 : HCCL_DEBUG("~TransportRemoteAccess Success!");
58 0 : }
59 0 : HcclResult TransportRemoteAccess::Init()
60 : {
61 0 : HCCL_DEBUG("TransportRemoteAccess Init start");
62 : // 创建QP操作句柄
63 0 : CHK_RET(CreateQp());
64 : // 本端host/device内存地址注册
65 0 : CHK_RET(MrRegister());
66 : // notify注册，用于rdma消息同步
67 0 : CHK_RET(NotifyRegister());
68 0 : CHK_RET(ConnectQp());
69 :
70 0 : HCCL_DEBUG("TransportRemoteAccess Init end");
71 0 : return HCCL_SUCCESS;
72 : }
73 :
74 0 : HcclResult TransportRemoteAccess::CreateQp()
75 : {
76 : // 创建qp handle, mode:普通qp
77 0 : HcclResult ret = HrtRaQpCreate(RemoteAccessPara_.nicRdmaHandle, QP_FLAG_RC, NORMAL_QP_MODE, handle_);
78 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Create][Qp]create qp mode failed, handle is null, "\
79 : "localRank[%u], qpMode[%d]", RemoteAccessPara_.localRank, NORMAL_QP_MODE), HCCL_E_ROCE_CONNECT);
80 0 : CHK_RET(SetQpAttrQos(handle_));
81 : // 配置RDMA Timeout时间
82 0 : CHK_RET(SetQpAttrTimeOut(handle_));
83 : // 配置RDMA Retry Cnt重传次数
84 0 : CHK_RET(SetQpAttrRetryCnt(handle_));
85 0 : return HCCL_SUCCESS;
86 : }
87 :
88 0 : HcclResult TransportRemoteAccess::NotifyRegister()
89 : {
90 : // notify内存注册
91 0 : CHK_RET(SetLocalNotify());
92 : // 获取对端notify
93 0 : CHK_RET(GetRemoteNotifyInfo());
94 0 : return HCCL_SUCCESS;
95 : }
96 :
97 0 : HcclResult TransportRemoteAccess::MrRegister()
98 : {
99 0 : void* memPtr = nullptr;
100 0 : if (MemRegistInfos_.size() == 0) {
101 0 : HCCL_ERROR("[Register][Mr]local mem info to register is empty!");
102 0 : return HCCL_E_PARA;
103 : }
104 0 : struct MrInfoT mrInfo = {nullptr};
105 0 : mrInfo.access = access_;
106 0 : for (size_t idx = 0; idx < MemRegistInfos_.size(); idx++) {
107 0 : memPtr = reinterpret_cast<void *>(static_cast<uintptr_t>(MemRegistInfos_[idx].addr));
108 0 : mrInfo.addr = memPtr;
109 0 : mrInfo.size = MemRegistInfos_[idx].length;
110 :
111 0 : CHK_RET(HrtRaMrReg(handle_, &mrInfo));
112 0 : localRegMem_.push_back(memPtr);
113 : }
114 0 : return HCCL_SUCCESS;
115 : }
116 :
117 0 : HcclResult TransportRemoteAccess::SetLocalNotify()
118 : {
119 : // 申请ack notify，并发送至对端
120 0 : CHK_RET(CreateNotify());
121 : // 注册notify 内存信息
122 0 : CHK_RET(CreateNotifyValueBuffer());
123 0 : return HCCL_SUCCESS;
124 : }
125 :
126 0 : HcclResult TransportRemoteAccess::CreateNotify()
127 : {
128 0 : u64 offset = 0;
129 0 : u64 notifyBaseVa = 0; // notify寄存器虚拟地址
130 0 : u64 notifyTotalSize = 0;
131 :
132 : /* 申请Notify Group ID */
133 0 : RemoteRankInfo info(deviceLogicId_, RemoteAccessPara_.remoteRank);
134 0 : CHK_RET(SalGetBareTgid(&info.remotePid)); // 当前进程id
135 0 : CHK_RET(notifyPool_->Alloc(tag_, info, ackNotify_));
136 : // 设置remote id
137 0 : s64 recvId = 0xFFFFFFFF00000000 | (static_cast<s64>(info.remotePid) & 0xFFFFFFFF);
138 0 : CHK_RET(ackNotify_->Grant(recvId));
139 :
140 : /* 获取notify寄存器虚拟基地址、大小, 物理地址回传值为空 */
141 0 : CHK_RET(HrtRaGetNotifyBaseAddr(RemoteAccessPara_.nicRdmaHandle, ¬ifyBaseVa, ¬ifyTotalSize));
142 :
143 : /* 获取notify虚拟地址 */
144 0 : CHK_RET(ackNotify_->GetNotifyOffset(offset));
145 :
146 : // notify寄存器的虚拟地址与物理地址偏移相同，所以虚拟地址为虚拟基地址加偏移
147 0 : u64 notifyVa = notifyBaseVa + offset;
148 :
149 0 : HCCL_INFO(
150 : "notifyBaseVa=0x%llx, notifyTotalSize=0x%x, offset=0x%llx, notifyVa=0x%llx ",
151 : notifyBaseVa, notifyTotalSize, offset, notifyVa);
152 :
153 : /* notify地址注册为mr, 在roce驱动中注册 */
154 0 : ackNotifyMsg_.mrRegFlag = 0; // mem注册给网卡标志位
155 0 : ackNotifyMsg_.addr = reinterpret_cast<void *>(static_cast<uintptr_t>(notifyVa)); // 本端notify地址交换给对端
156 0 : ackNotifyMsg_.len = notifySize_;
157 0 : ackNotifyMsg_.offset = offset;
158 :
159 0 : return HCCL_SUCCESS;
160 : }
161 :
162 0 : HcclResult TransportRemoteAccess::GetRemoteNotifyInfo()
163 : {
164 0 : NotifyMsg mrMsg;
165 0 : s32 sRet = memset_s(&mrMsg, sizeof(NotifyMsg), 0, sizeof(NotifyMsg));
166 0 : CHK_PRT_RET(sRet != EOK, HCCL_ERROR("[Get][NotifyInfo]errNo[0x%016llx]get remote addr, memory set 0 failed. "\
167 : "params: destMaxSize[%zu], count[%zu]", HCCL_ERROR_CODE(HCCL_E_MEMORY), \
168 : sizeof(NotifyMsg), sizeof(NotifyMsg)), HCCL_E_MEMORY);
169 :
170 0 : CHK_RET(hrtRaSocketBlockRecv(RemoteAccessPara_.socketFdhandle, &mrMsg, sizeof(NotifyMsg)));
171 0 : sRet = memcpy_s(&remoteNotifyDataMsg_, sizeof(NotifyMsg), &mrMsg, sizeof(NotifyMsg));
172 0 : CHK_PRT_RET(sRet != EOK, HCCL_ERROR(
173 : "[Get][NotifyInfo]errNo[0x%016llx] In TransportRemoteAccess get remote addr, memcpy failed. "\
174 : "errorno[%d], params:destMaxSize[%zu],count[%zu]", HCCL_ERROR_CODE(HCCL_E_MEMORY),
175 : sRet, sizeof(NotifyMsg), sizeof(NotifyMsg)), HCCL_E_MEMORY);
176 0 : HCCL_INFO("recv success:len=%llu", mrMsg.len);
177 :
178 0 : return HCCL_SUCCESS;
179 : }
180 :
181 0 : HcclResult TransportRemoteAccess::CreateNotifyValueBuffer()
182 : {
183 0 : std::unique_lock<std::mutex> lock(notifyValueMutex_[deviceLogicId_]);
184 0 : if (notifyValueMem_[deviceLogicId_].ptr() == nullptr) {
185 0 : u64 notifyVaule = 1; // notify值写1表示record
186 0 : CHK_RET(DeviceMem::alloc(notifyValueMem_[deviceLogicId_], notifyValueSize_));
187 0 : HCCL_DEBUG("create notify value size[%u]", notifySize_);
188 :
189 0 : CHK_RET(hrtMemSyncCopy(notifyValueMem_[deviceLogicId_].ptr(), notifyValueMem_[deviceLogicId_].size(),
190 : ¬ifyVaule, notifySize_, HcclRtMemcpyKind::HCCL_RT_MEMCPY_KIND_HOST_TO_DEVICE));
191 : }
192 0 : lock.unlock();
193 0 : struct MrInfoT mrInfo = {nullptr};
194 0 : mrInfo.addr = notifyValueMem_[deviceLogicId_].ptr();
195 0 : mrInfo.size = notifySize_;
196 0 : mrInfo.access = access_;
197 0 : CHK_RET(HrtRaMrReg(handle_, &mrInfo));
198 : // 将notify buffer地址保存
199 0 : localRegMem_.push_back(notifyValueMem_[deviceLogicId_].ptr());
200 0 : NotifyMsg msg = {};
201 0 : msg.mrRegFlag = REG_VALID;
202 0 : msg.addr = notifyValueMem_[deviceLogicId_].ptr();
203 0 : msg.len = notifySize_;
204 :
205 : /* 发送mem消息给对端 */
206 0 : HcclResult ret = hrtRaSocketBlockSend(RemoteAccessPara_.socketFdhandle, &msg, sizeof(NotifyMsg));
207 0 : if (ret != HCCL_SUCCESS) { // 发送成功字节数与发送字节数不等，发送失败
208 0 : HCCL_ERROR("[Create][NotifyValueBuffer]send=%zu", sizeof(NotifyMsg));
209 0 : return HCCL_E_INTERNAL;
210 : }
211 :
212 0 : return HCCL_SUCCESS;
213 0 : }
214 0 : HcclResult TransportRemoteAccess::RemoteRead(const std::vector<HcomRemoteAccessAddrInfo>& addrInfos, Stream &stream)
215 : {
216 0 : CHK_SMART_PTR_NULL(stream);
217 0 : HCCL_INFO("TransportRemoteAccess RemoteRead begin");
218 0 : CHK_RET(RdmaDataTransport(addrInfos, RDMA_OP_READ));
219 :
220 : // 读取远端notify buffer，用于notify同步
221 0 : CHK_RET(ReadRemoteNotifyBuffer());
222 :
223 : // 等待TS把任务处理完成
224 0 : CHK_RET(LocalIpcNotify::Wait(stream, const_cast<HcclDispatcher>(dispatcher_), ackNotify_, INVALID_VALUE_STAGE,
225 : NOTIFY_INVALID_WAIT_TIME, RemoteAccessPara_.localRank, RemoteAccessPara_.remoteRank));
226 :
227 0 : HCCL_INFO("TransportRemoteAccess RemoteRead end");
228 0 : return HCCL_SUCCESS;
229 : }
230 :
231 0 : HcclResult TransportRemoteAccess::RemoteWrite(const std::vector<HcomRemoteAccessAddrInfo>& addrInfos, Stream &stream)
232 : {
233 0 : CHK_SMART_PTR_NULL(stream);
234 0 : HCCL_RUN_INFO("TransportRemoteAccess RemoteWrite begin, addressNum[%u]", addrInfos.size());
235 0 : CHK_RET(RdmaDataTransport(addrInfos, RDMA_OP_WRITE));
236 : // 读取远端notify buffer，用于notify同步
237 0 : CHK_RET(ReadRemoteNotifyBuffer());
238 : // 等待TS把任务处理完成
239 0 : CHK_RET(LocalIpcNotify::Wait(stream, const_cast<HcclDispatcher>(dispatcher_), ackNotify_, INVALID_VALUE_STAGE,
240 : NOTIFY_INVALID_WAIT_TIME, RemoteAccessPara_.localRank, RemoteAccessPara_.remoteRank));
241 :
242 0 : return HCCL_SUCCESS;
243 : }
244 :
245 0 : HcclResult TransportRemoteAccess::RdmaDataTransport(const std::vector<HcomRemoteAccessAddrInfo>& addrInfos, s32 rdmaOp)
246 : {
247 0 : if ((rdmaOp != RDMA_OP_WRITE) && (rdmaOp != RDMA_OP_READ)) {
248 0 : HCCL_ERROR("[Transport][RdmaData]invalid rdma op type, op:[%d]", rdmaOp);
249 0 : return HCCL_E_PARA;
250 : }
251 0 : CHK_PRT_RET(addrInfos.empty(), HCCL_ERROR("[Transport][RdmaData]addrInfos is empty!"), HCCL_E_PARA);
252 0 : u32 addressNum = addrInfos.size();
253 : // 构造wr信息
254 0 : std::vector<struct SendWrlistDataExt> wrVec(addressNum);
255 0 : std::vector<struct SendWrRsp> opRspVec(addressNum);
256 0 : struct SendWrlistDataExt *wr = wrVec.data();
257 0 : struct SendWrRsp *opRsp = opRspVec.data();
258 0 : struct SgList list = {0};
259 0 : u64 length = addrInfos[0].length;
260 :
261 0 : HCCL_RUN_INFO("RdmaDataTransport begin, addressNum[%u], length[%u]", addressNum, length);
262 0 : for (size_t idx = 0; idx < addrInfos.size(); idx++) {
263 0 : list.addr = static_cast<u64>(static_cast<uintptr_t>(addrInfos[idx].localAddr));
264 0 : list.len = addrInfos[idx].length;
265 :
266 0 : wr[idx].memList = list;
267 0 : wr[idx].dstAddr = static_cast<u64>(static_cast<uintptr_t>(addrInfos[idx].remoteAddr));
268 0 : wr[idx].op = rdmaOp; /* RDMA_WRITE: 0 RDMA_READ: 4 */
269 0 : wr[idx].sendFlags = RA_SEND_SIGNALED;
270 : }
271 0 : u32 nowIdex = 0;
272 0 : u32 singleCompleteNum = 0;
273 0 : u32 sendNum = 0;
274 : HcclResult ret;
275 0 : u32 tryCount = 0;
276 0 : while (nowIdex < addressNum) {
277 0 : sendNum = addressNum - nowIdex;
278 0 : ret = HrtRaSendWrlistExt(handle_, wr, opRsp, sendNum, &singleCompleteNum);
279 0 : if (ret == HCCL_SUCCESS) {
280 0 : HCCL_INFO("dlRaSendWrlist success singleCompleteNum[%u], addressNum[%u]", singleCompleteNum, addressNum);
281 0 : return HCCL_SUCCESS;
282 0 : } else if (ret == ENOENT) { // 未完成发送，需重试
283 0 : nowIdex += singleCompleteNum;
284 0 : wr += singleCompleteNum;
285 0 : opRsp += singleCompleteNum;
286 0 : tryCount++;
287 0 : CHK_PRT_RET(tryCount > SEND_WRLIST_MAX_COUNT,
288 : HCCL_ERROR("[Transport][RdmaData]dlRaSendWrlist count beyond maxnum[%u], completenum[%u]",
289 : SEND_WRLIST_MAX_COUNT, nowIdex), HCCL_E_NETWORK);
290 0 : continue;
291 : } else {
292 0 : HCCL_ERROR("[Transport][RdmaData]In RdmaDataTransport, hrtRaSendWrlist failed. op[%d], ret[%d]",
293 : rdmaOp, ret);
294 0 : return HCCL_E_NETWORK;
295 : }
296 : }
297 :
298 0 : return HCCL_SUCCESS;
299 0 : }
300 :
301 0 : HcclResult TransportRemoteAccess::ReadRemoteNotifyBuffer()
302 : {
303 0 : HCCL_INFO("In TransportRemoteAccess ReadRemoteNotifyBuffer begin");
304 0 : struct SgList list = {0};
305 0 : struct SendWr wr = {nullptr};
306 :
307 : // 构造wr信息
308 0 : list.addr = static_cast<u64>(reinterpret_cast<uintptr_t>(ackNotifyMsg_.addr));
309 0 : list.len = remoteNotifyDataMsg_.len;
310 :
311 0 : wr.bufList = &list;
312 0 : wr.bufNum = 1; /* 此处list只有一个，设置为1 */
313 0 : wr.dstAddr = static_cast<u64>(reinterpret_cast<uintptr_t>(remoteNotifyDataMsg_.addr));
314 0 : wr.op = RDMA_OP_READ; /* RDMA_WRITE: 0 */
315 0 : wr.sendFlag = RA_SEND_SIGNALED;
316 :
317 0 : struct SendWrRsp opRsp = {0};
318 0 : CHK_RET(HrtRaSendWr(handle_, &wr, &opRsp));
319 0 : HCCL_INFO("In TransportRemoteAccess ReadRemoteNotifyBuffer end");
320 0 : return HCCL_SUCCESS;
321 : }
322 :
323 0 : HcclResult TransportRemoteAccess::ConnectQp()
324 : {
325 : // QP建链
326 0 : CHK_RET(HrtRaQpConnectAsync(handle_, RemoteAccessPara_.socketFdhandle));
327 :
328 0 : HCCL_INFO("TransportRemoteAccess ConnectQp LocalRank[%u] "\
329 : "RemoteRank[%u] LocalIp[%s]", RemoteAccessPara_.localRank,
330 : RemoteAccessPara_.remoteRank, RemoteAccessPara_.localIp.GetReadableAddress());
331 :
332 : // 查询QP建链是否成功
333 0 : s32 qpStatus = 0;
334 0 : auto startTime = std::chrono::steady_clock::now();
335 0 : HCCL_INFO("In link ibv, waiting for qp status ready...");
336 : while (true) {
337 0 : if ((std::chrono::steady_clock::now() - startTime) >= timeout_) {
338 0 : HCCL_ERROR("[Connect][Qp]get qp status timeout_=%lld s, qp_status=%d", timeout_, qpStatus);
339 0 : return HCCL_E_TIMEOUT;
340 : }
341 :
342 0 : s32 raRet = hrtGetRaQpStatus(handle_, &qpStatus);
343 0 : if ((!raRet) && (qpStatus == 1)) { // 为1时，qp 建链成功
344 0 : HCCL_INFO("GetRaQpStatus Success!");
345 0 : break;
346 : } else {
347 : // qp建链需要时间，获取qp状态直至超时
348 0 : SaluSleep(WAIT_US_COUNT);
349 : }
350 0 : }
351 0 : return HCCL_SUCCESS;
352 : }
353 : }
|
